Content uploaded by Mika Hämäläinen
Author content
All content in this area was uploaded by Mika Hämäläinen on Dec 07, 2023
Content may be subject to copyright.
Working Towards Digital Documentation of Uralic Languages With
Open-Source Tools and Modern NLP Methods
Mika Hämäläinen1, Jack Rueter 2, Khalid Alnajjar3and Niko Partanen2
1Metropolia University of Applied Sciences
2University of Helsinki
3Rootroo Ltd
firstname.lastname@metropolia.fi/helsinki.fi/rootroo.com
Abstract
We present our work towards building an in-
frastructure for documenting endangered lan-
guages with the focus on Uralic languages in
particular. Our infrastructure consists of tools
to write dictionaries so that entries are struc-
tured in XML format. These dictionaries are
the foundation for rule-based NLP tools such
as FSTs. We also work actively towards en-
hancing these dictionaries and tools by using
the latest state-of-the-art neural models by gen-
erating training data through rules and lexica.
1 Introduction
Most of the languages spoken in the world are in
danger of extinction. Their documentation and revi-
talization are of a highest cultural value, for which
they have received plenty of academic attention
in various disciplines such as anthropology, typol-
ogy, lexicography and computational linguistics.
Needless to say, the resources produced in each in-
dividual research project are not always published
openly let alone made available to the community
of native speakers.
The goal of our paper is to describe our open
infrastructure for documenting minority languages.
We present our experiences with the following
Uralic languages: Skolt Sami (sms), Erzya (myv),
Moksha (mdf), Komi-Zyrian (kpv) and Komi-
Permyak (koi). As they belong to the Uralic branch,
they are languages that exhibit a complex morphol-
ogy, which makes their computational processing
a challenge for modern machine learning methods
that would require a lot of data to cover this com-
plexity. The quantity and quality of data is usually
an issue when we deal with endangered languages
(Hämäläinen,2021). Carrying out linguistic docu-
mentation in a structured machine readable format,
however, makes it possible to create the resources
needed for building NLP tools simultaneously with
linguistic documentation.
We are about to start working with the Apurinã
(apu) language, which allows us to reflect upon
our Uralic context from a broader perspective, and
increases the relevance of our work in a Latin Amer-
ican context. Thus, we describe how our infrastruc-
ture can work in non-Uralic contexts.
Linguistic documentation is a field of academic
study that has developed considerably in recent
decades. Its purpose is to provide a complete record
of the linguistic practices characteristic to a given
speech community (Himmelmann,1998). The goal
of language documentation is to describe the lan-
guage of a community of speakers as fully as pos-
sible both for future generations and for language
revitalization. The result of this work is typically
manifested as a linguistic corpus or other type of
material collection, which later on can be stud-
ied and analyzed in various ways. The question
whether the collected materials actually describe
the language use of a speech community is de-
batable, and this goal can never be fully achieved
because a corpus can never describe a language in
full. Nonetheless, linguistically collected materials
may be the only resources available for a small
language.
Whether and how language documentation mate-
rials should be made accessible and distributed, has
been a matter of debate. We believe it is important
to understand that this is also a matter of granular-
ity, and the question is not necessarily whether the
materials are accessible, but rather which parties
should be allowed what type of access. There are
good reasons for keeping culturally sensitive mate-
rials available only to specific groups. At the same
time, there are always materials in any language
that are more neutral and such that the authors them-
selves may want to make accessible. Especially for
written publications, it may always be possible to
negotiate a publication with open licenses, which
would also allow the reuse of the same materials in
different open research purposes.
Open materials are particularly important when
we develop tools for NLP, because this work can
greatly benefit from resources that are openly ac-
cessible with a permissive license. In the following
sections we will discuss examples of such work, in-
cluding our contribution to Universal Dependency
treebanks. It must be emphasized that the open
technology developed on an open infrastructure can
also be used to process materials that are available
only to a particular researcher or individual mem-
bers of a community. Therefore, open infrastruc-
ture benefits both open and closed environments,
whereas a closed infrastructure only benefits a big
commercial player.
2 Related work
There are several individual projects in different
parts of the world that work with online dictio-
naries for endangered languages. Many projects,
however, focus on one language only and work
without knowing about other ongoing projects for
other endangered languages. This has led to a sit-
uation where researchers solve the same type of
problems individually for their language of interest
reinventing the wheel over and over again. There
are plenty of online dictionaries and language learn-
ing tools that have been developed from scratch for
one particular language.
Work with endangered languages in North Amer-
ica has shown the importance of language learning
tools for second language learners. Lack of famil-
iarity with lexicographical tradition can easily be a
deciding factor in a beginner’s learning experience.
A learner of a new language cannot be expected to
know exactly where an entry is located in a dictio-
nary, nor can the learner be expected automatically
to know the normative spelling. When the user
of a language lacks a proper keyboard layout or
knowledge of the correct orthography, the strate-
gies of orthographic relaxation can be implemented
in mobile and online dictionaries. Morphological
processing and spelling relaxation are used to cater
to beginners in Tsimshian and Salishan languages
in the use of dictionaries and NLP tools (Littell
et al.,2017).
On an entirely separate front, work has also
been done to provide the Yupik community of St.
Lawrence Island unimpeded access to language
materials online. This has been possible using a
morphologically aware dictionary. In the system, a
strategy of multiple input methods has been intro-
duced that caters to different writing systems (Hunt
et al.,2019). The work here is tailor-made, and it
maintains a strong link between the language and
its community. The endangered language is seen
as a low-resourced language in this context.
The problem is that low-resourced language is
a term that is used for almost any language with
less Internet presence than English. languages
like Hindi (Irvine and Callison-Burch,2014), Ara-
bic (Chen et al.,2018) or Persian (Ahmadnia
et al.,2017) are often considered low-resourced
languages in the world of NLP, even though they
have millions of speakers. In the work of Na-
sution et al. (2018), the ethnic Indonesian lan-
guages are relatively small compared to the su-
perstrate language that surrounds them. The ap-
proach consists of working simultaneously with a
group of closely related languages in a multilingual,
language-independent infrastructure. The authors
analyze the use of bilingual dictionary entries and
explain the difficulty of selecting the appropriate
bilingual dictionaries to begin documentation.
One of the largest infrastructures for minority
language documentation from the point of view of
computational linguistics is that of Giella (Mosha-
gen et al.,2014). Their infrastructure is based on
two main components: FST transducers (finite state
transducers) and XML dictionaries. Transducers
are a way of documenting the morphology of a
language computationally. That is to say, they are
collections of rules about how the morphological
system of a language works. These rules can be
used directly for automatic text analysis and lemma
conjugation in its morphological variants.
Transducers and XML dictionaries are used for
spelling correction in Word
1
, text prediction on An-
droid and iOS keyboards
2
, interactive systems to
learn languages (Bontogon et al.,2018) and online
dictionaries (Rueter and Hämäläinen,2017). Our
infrastructure is based on Giella, which allows us to
synchronize data between the two infrastructures.
This means that advances in linguistic documenta-
tion in our infrastructure can be used directly in the
tools produced in Giella.
3 Our infrastructure
Using Giella requires a relatively high proficiency
in programming to be able to write dictionaries and
morphological rules for FSTs, and at the same time,
1http://divvun.no/korrektur/korrektur.html
2http://divvun.no/keyboards/mobileindex.html
Figure 1: The form in Akusanat to edit the entry piânnai (dog) in Skolt Sami
it requires a good amount of knowledge in the lan-
guage that is being documented. The infrastructure
can be too complicated even for those who have
studied computer science, and therefore it is not
accessible to a community outside of those who col-
laborate directly with Giella. For this reason, our
infrastructure has several interfaces for different
types of users; for users who do not have sufficient
knowledge to write XML or program transducers
and for developers who want to use the tools with-
out knowing how to compile them right from the
beginning with the make command.
3.1 Online dictionaries
A very important step in the documentation of a mi-
nority language is the lexicographical work. This
results in a dictionary that can be useful for both
native speakers as for those who want to learn the
language. We store dictionaries in a highly struc-
tured XML format. That means that all kinds of
metadata are in their respective fields rather than
being stored in various parts of a lexicographical
entry in an unstructured format. This is important
as we do not only want to create dictionaries for
human use, but we also want them to be machine
readable.
Our Akusanat system
3
(Hämäläinen and Rueter,
2018,2019) is based on MediaWiki and allows you
to view the content of XML dictionaries for all
types of users. MediaWiki data is synchronized
with XML files using the Git version control. This
means that if someone modifies a lexicographical
entry in Akusanat, these changes will result in a
change to the XML dictionary stored in GitHub. If
3https://akusanat.com.
someone changes the XML dictionaries directly,
Akusanat will download the new changes from
GitHub and update its database automatically. This
is done so that advanced users are able to edit the
XML files directly with their favorite tool and less
advanced users can make changes online with a
graphical user interface. Akusanat does not let
users modify the Wiki syntax directly, instead it
displays a form that ensures changes remain struc-
tured and compatible with XML Figure 1.
For searching, we use morphological FST trans-
ducers to process the user input. This means that
the user can search for a word in any of its morpho-
logical inflections, since the FST can lemmatize
words automatically. It is also possible to search by
typing in misspelled words. The transducers con-
tain information about the most common spelling
errors in each language, which allows us to resolve
the lemma, although the word has not been spelled
according to the spelling standard. This is impor-
tant in the case of languages with which we work,
since spelling rules are not as well-established as
in the case of majority languages.
Figure 3shows the interface for looking up
words in the dictionaries. In the example, the
search term is the Skolt Sami word soo
gg
, which is
the genitive of the word sokk, which means family.
Our system lemmatizes the search term automat-
ically with the Skolt Sami FST, and displays the
input for the sokk lemma to the user.
The idea of using MediaWiki, and especially Se-
mantic MediaWiki, to create dictionaries, is not
new, since there are already several projects that
use the technology as their base (Muljadi et al.,
2006;Bon and Nowak,2013;Dueñas and Gómez,
Figure 2: The interface for searching and filtering lexical entries in Ve’rdd
Figure 3: The interface for searching in Akusanat
2015). Without a doubt, MediaWiki has its ad-
vantages, in practice we have had to program our
own MediaWiki extensions to add the necessary
functionality; the form to edit, MediaWiki-XML
synchronization, search with transducers etc. The
problem that we have experienced many times is
that the inner workings of MediaWiki change too
often. This means that if we want to keep our Me-
diaWiki instance up to date with the latest security
updates, we have to make a lot of changes to our
source code to keep our extensions working with
the new version of MediaWiki. Even so, we con-
tinue to use and develop Akusanat
4
for the time
being, as it offers a simple environment for users.
In the next section, we describe the other system
4Code available https://github.com/mikahama/akusanat
that we are developing. The new system may re-
place Akusanat in the future.
3.2 Editorial work
In this section, we describe the Ve’rdd
5
system (Al-
najjar et al.,2019,2020). The system works with
the same XML dictionaries as Akusanat and can be
used online in a similar way. The difference is in
the intended use of the system. Ve’rdd is not a sys-
tem to visualize lexicographical entries for an end
user, but a system created specifically for writing
both digital and printed dictionaries. During the
process of developing the system, we have collabo-
rated with a group of professional lexicographers
who work with printed dictionaries.
In the context of the languages we work with,
lexicographical documentation does not start from
scratch, as both the Sami languages spoken in the
Nordic countries and the Permian and Mordvinic
languages spoken in Russia have received much
attention in terms of their digital documentation
during the last century. For example, there is a
dictionary of the Skolt Sami language Sammallahti
and Mosnikoff (1991), and there are several stud-
ies on the Mordvinic (Aasmäe et al.,2016;Grün-
thal,2016) and Permian languages (Hamari,2011;
Klumpp,2016). If there are existing dictionaries in
digital form, they exist in an unstructured format
such as a Word, CSV, or PDF file produced with
an OCR system. For this reason, Ve’rdd includes
functionality for import lexicographic data from un-
structured formats. We have paid a lot of attention
5https://akusanat.com/verdd/
Figure 4: The interface for editing lexical entries in Ve’rdd
to the quality of the conversion, since, in the case of
our languages, especially in the case of Skolt Sami,
it is very frequent that the same character exists in
many different Unicode characters. For example,
′
(U+02B9 modifier letter prime) is a very common
character in Skolt Sami, but because of the Finnish
keyboard layout, it is often written as ’ (U+0027
apostrophe) or
´
(U+00B4 acute accent). Ve’rdd
is programmed to take into account the possible
characters of the language and try to correct the
incorrect characters automatically.
Figure 2shows the interface for searching and
filtering words in Ve’rdd. The interface is designed
to support the workflow of a dictionary editor. For
example, it is possible to display only raw inputs.
This means entries that no one has verified after
importing the data from an unstructured format. To
facilitate the development of FST transducers it is
also possible to sort and filter the words according
to the continuation lexicon, which is the FST way
of indicating how every word is supposed to be
inflected.
Apart from just searching and filtering lexical
entries, it is important to have the possibility to edit
them. Figure 4shows the interface for inspecting
a dictionary entry. If a user is connected to their
account, in addition to viewing, they can edit the
information of a lexicographic entry. Ve’rdd is de-
signed to be a tool for multilingual dictionaries, so
one entry is connected to other entries in the system.
In Figure 4, relationships can be seen as translation
types that connect a word to its translations in other
languages. It is also possible to define other types
of relationships between lexica based on etymol-
ogy. Relationships may also exist between words
of the same language, for example, it is possible
to indicate compound words and derivations with
relations. Since the FST transducers contain deriva-
tive information, Ve’rdd automatically adds this
type of relationship when importing a unstructured
dictionary.
Ve’rdd can visualize the relationship between
two words that are linked together with any kind
of relationship. This can be used to verify that a
word in a given language is linked to the correct
homonym in another language (Figure 5). It is also
possible to edit the type of relationship or delete
any unnecessary relationships.
Ve’rdd has a functionality that allows the user
to export any dictionary in different formats. The
most important for us are the Giella XML, which
can be used to generate FST transducers, and Latex
Figure 5: The interface for comparing two related entries in Ve’rdd
code. The Latex code makes it is possible to gener-
ate a ready-to-print PDF version of the dictionary.
The Latex format makes it possible to change the
style of the dictionary without changing the con-
tent. If there are changes in Ve’rdd, it is possible to
update the content of the dictionary without chang-
ing the style defined in Latex. This functionality
has been an important design principle for us since
the work done in Ve’rdd should not only be used in
digital dictionaries but also in printed dictionaries.
3.2.1 NLP resources
Our dictionary editing systems are directly useful
in the development of FST transducers since we
can export the lexicon in the format needed for
HFST (Lindén et al.,2013). HFST is the tool we
use to create the transducers. We have transduc-
ers for the Skolt Sami (Rueter and Hämäläinen,
2020), Erzya and Moksha (Rueter et al.,2020a)
and Komi languages. The transducers can be used
to lemmatize words, analyze their morphology or
generate inflected forms. These transducers are dif-
ficult to compile for people who do not work with
the transducers often. For this reason, we compile
all transducers every night and we distribute them
through our website
6
. We not only compile our
transducers but all transducers for all languages in
the Giella infrastructure.
The transducers are difficult to use as such, and
for this reason, we have developed a Python li-
brary called UralicNLP (Hämäläinen,2019) and a
Python implementation of HFST called PyHFST
(Hämäläinen and Alnajjar,2023). With the li-
braries, compiled dictionaries and translators can
be downloaded and used directly in Python. Fig 6
shows how to use our transducers from Python. In
the second line of code, the word шляпа (hat) is
analyzed in erzya (myv). The result indicates that
the word is an indefinite (+Indef) noun (+N) in the
nominative (+Nom) singular (+Sg). In the fourth
line we generate the conjugated form of the same
word in the plural (+Pl). The result is the plural
word шляпат.
Figure 6: An example of using UralicNLP
6https://models.uralicnlp.com/nightly/
FST transducers produce all possible interpre-
tations for a word from. In the case of the Uralic
languages, there is plenty of homonymy in morpho-
logical inflections. This means that, if we use the
transducers on regular text, we cannot accurately
lemmatize the words in their context since the trans-
ducers produce all possible lemmas, For this reason,
we use constraint grammar disambiguators (Karls-
son et al.,2011) based on a tool called VISL CG-3
(Bick and Didriksen,2015). The grammar rules of
constraint grammars remove morphological read-
ings that are not possible in a given sentence, and
result in a sentence that is morphologically disam-
biguated with one lemma per word as opposed to
all the possible lemmas.
Figure 7: An example of the use of the Komi Zyrian
disambiguator
In Figure 7, we can see how the CG disambigua-
tors can be used on UralicNLP. The third line initial-
izes the disambiguation object for the Komi-Zyrian
(kpv) and in the fourth line the disambiguation
method of the object is called with a sentence. The
result contains the word forms of the sentence, their
lemmatization and morphology for each word of
the sentence.
Apart from structured dictionaries and rule-
based tools, we have treebanks of the universal
dependencies for the Skolt Saami, Moksha, Erzya
(Rueter and Tyers,2018), Komi-Zyrian (Partanen
et al.,2018) and Komi-Permyak (Rueter et al.,
2020b). These treebanks contain syntactic annota-
tions with the tags Morphological characteristics of
universal dependencies. With the latest treebanks,
we have also added the morphological labels pro-
duced by the transducers to facilitate the use of the
two resources together
4 Incorporating modern NLP methods
As we have described thus far, a great part of our
work relies on the old rule-based tradition of NLP.
When we deal with endangered languages, rules are
the primary starting point. One cannot simply train
a neural network if there isn’t enough training data.
However, we do not want to reject neural models
instantly as something that simply will not work
for small languages. Neural models can work and
they can be extremely beneficial. Throughout our
research, we have aimed at combining rule-based
models with neural models to facilitate our work
on endangered languages.
Digital documentation has allowed us to use the
latest methods in the world of NLP to automati-
cally increase the data we have in the dictionaries.
Because all of the lexicographic resources we have
are multilingual, the first step we have taken with
NLP technology has been the prediction of trans-
lations (Hämäläinen et al.,2018). The idea was
as follows: if the Skolt Sami dictionary contains
Finnish translations, German and English, and the
Erzya dictionary contains translations into Finnish,
English, Russian, and French, then, with this in-
formation, it should be possible to automatically
deduce translations from Skolt Sami into Russian
and French and from Erzya into German given the
existence of two common languages: Finnish and
English. With a probabilistic model we have in-
creased the number of translations in Skolt Sami,
Erzya, Moksha and Komi-Zyrian dictionaries.
We have elaborated on this idea later on by us-
ing graph based approaches and neural models (Al-
najjar et al.,2021,2022). These have not been
isolated attempts, but the graph based methods
have been incorporated into Ve’rdd as well. The
predictions have been manually checked and this
way we have been able to augment our dictionaries
semi-automatically. The Livonian institute has em-
braced this technology in bootstrapping a Livonian-
English dictionary.
As neural networks require a large amount of
data to be trained, it is common to believe that their
use is not possible in the case of endangered lan-
guages. We have taken the perspective that we can
generate the amount of data needed for a neural
network with our morphological tools. Using the
treebanks and the transducers, we have generated
data to train a neural network to perform disam-
biguation instead of using the constraint grammar
for Erzya and Komi-Zyrian (Ens et al.,2019). The
idea was to generate all possible analyzes for the
words in the treebanks and train the neural network
to disambiguate the analyses with the treebank anal-
ysis. Later on, we further developed this method
in the context of Sami languages (Hämäläinen and
Wiechetek,2020).
We have also been able to use the neural net-
works to increase etymological relationships in
the Skolt Sami dictionary (Hämäläinen and Reuter,
2019). The method was based on a character level
LSTM model that was enhanced with synthetic data
generated with a character-level statistical machine
translation tool. We used this method to produce a
set of candidate cognates that we manually checked
and incorporated into our digital dictionaries. This
method relies on external data from the Institute
for Languages in Finland, which makes it difficult
for us to include it in Ve’rdd.
Rule-based FSTs are great because they are usu-
ally very accurate, however, they do not have a
great lexical coverage. Analyzing an online text
with the FSTs will usually mean a ton of words that
are not recognized at all. For this reason, we used
the FSTs to generate training data for neural mod-
els (Hämäläinen et al.,2021). We used this data
to train character-level neural machine translation
models to analyze, generate and lemmatize word
forms. The key idea is to use the exact same mor-
phological tags so that the neural models and the
FSTs can be used interchangeably. These neural
models have been made available through Uralic-
NLP as a fallback mechanism. If an FST fails to
analyze a word form, the neural model will be used
automatically if neural fallback is turned on.
Recently, we have also moved our interest to-
wards other aspects of NLP than just lexicon and
morphology. We have done work on automatically
translating and aligning word embeddings for en-
dangered Uralic languages (Alnajjar,2021) and
using them successfully in downstream tasks such
as sentiment analysis (Alnajjar et al.,2023).
5 Discussion and Conclusions
We hope that our work can be useful for others
as well. We have put a lot of attention in open-
sourcing our tools and resources so that nobody
needs to start building language documentation
tools entirely from scratch. We have also paved
a road towards using state-of-the-art neural models
in the context of truly endangered languages with
extremely limited resources. This is challenging
and requires ingenuity. We are not interested in
committing to the dichotomy of researches who
defend rule-based tools as the only viable option
for endangered languages nor to the researchers
who frown upon rules and rely solely on the Trans-
former architecture. The best solutions, we believe,
are found by combining both worlds.
Our tools are compatible with the Giella infras-
tructure. This has made it possible to use our dic-
tionaries and translators directly on their online
platform to learn languages (Antonsen and Argese,
2018), on Android and iPhone keyboards and spell
checking for Word and OpenOffice developed by
Divvun
7
at Giella. Flexible and interoperable de-
sign makes it also possible to integrate different
lexical resources into our infrastructure once those
are digitized or otherwise become available.
Digital documentation clearly has its benefits,
since we can carry out machine learning with struc-
tured dictionaries and FST transducers. For this
reason, a project conducted at the University of
Oulu
8
the goal of which was to author the new
dictionary Skolt Finnish-Sami has chosen to use
Ve’rdd to create the digital and printed dictionary.
We have worked together with project employees
to increase the functionality of our system. Ve’rdd
has made the simultaneous work of editors possi-
ble who, without Ve’rdd, would have used Excel
and Word. This would have meant a lost chance of
producing a structured dictionary for the interest of
NLP and a printed dictionary at the same time.
We have started to explore non-Uralic languages
by building a UD treebak for Apurinã (Rueter et al.,
2021). Furthermore, we have built an initial FST
for Lushootseed (lut) (Rueter et al.,2023) and ex-
tended it with an LSTM model. These are our
initial steps towards non-Uralic languages.
References
Niina Aasmäe, Karl Pajusalu, and NADEŽDA KABA-
JEVA. 2016. Gemination in the mordvin languages.
Linguistica Uralica, 52(2).
Benyamin Ahmadnia, Javier Serrano, and Gholamreza
Haffari. 2017. Persian-spanish low-resource statis-
tical machine translation through english as pivot
language. In RANLP, pages 24–30.
Khalid Alnajjar. 2021. When word embeddings become
endangered. In Multilingual Facilitation, pages 275–
288. University of Helsinki.
Khalid Alnajjar, Mika Hämäläinen, Niko Partanen, and
Jack Rueter. 2019. The open dictionary infrastructure
for uralic languages. Электронная Письменность
Народов Российской Федерации.
Khalid Alnajjar, Mika Hämäläinen, Niko Tapio Parta-
nen, and Jack Rueter. 2022. Using graph-based meth-
ods to augment online dictionaries of endangered
languages. In Proceedings of the Fifth Workshop on
7http://divvun.no/
8
https://www.sttinfo.fi/tiedote/tekoaly-apuna-
koltansaamen-ja-pohjoissaamen-digitaalisten-sanakirjojen-
toimitustyossa?publisherId=57858920&releaseId=69886820
the Use of Computational Methods in the Study of
Endangered Languages, pages 139–148.
Khalid Alnajjar, Mika Hämäläinen, and Jack Rueter.
2023. Sentiment analysis using aligned word em-
beddings for uralic languages. In Proceedings of
the Second Workshop on Resources and Representa-
tions for Under-Resourced Languages and Domains
(RESOURCEFUL-2023), pages 19–24.
Khalid Alnajjar, Mika Hämäläinen, Jack Rueter, and
Niko Partanen. 2020. Ve’rdd. narrowing the gap be-
tween paper dictionaries, low-resource nlp and com-
munity involvement. In Proceedings of the 28th Inter-
national Conference on Computational Linguistics:
System Demonstrations, pages 1–6.
Khalid Alnajjar, Jack Rueter, Niko Partanen, and Mika
Hämäläinen. 2021. Enhancing the erzya-moksha
dictionary automatically with link prediction. Folia
Uralica Debreceniensia.
Lene Antonsen and Chiara Argese. 2018. Using authen-
tic texts for grammar exercises for a minority lan-
guage. In Proceedings of the 7th workshop on NLP
for Computer Assisted Language Learning, pages
1–9.
Eckhard Bick and Tino Didriksen. 2015. Cg-3—beyond
classical constraint grammar. In Proceedings of the
20th Nordic Conference of Computational Linguistics
(NODALIDA 2015), pages 31–39.
Bruno Bon and Krzysztof Nowak. 2013. Wikilexico-
graphica. linking medieval latin dictionaries with se-
mantic mediawiki. In eLex 2013, pages 407–420.
Trojina, Institute for Applied Slovene Studies (Ljubl-
jana, Slovenia); Eesti . . . .
Megan Bontogon, Antti Arppe, Lene Antonsen, Dorothy
Thunder, and Jordan Lachler. 2018. Intelligent
computer assisted language learning (icall) for
nêhiyawêwin: an in-depth user-experience evaluation.
Canadian Modern Language Review, 74(3):337–362.
Xilun Chen, Yu Sun, Ben Athiwaratkun, Claire Cardie,
and Kilian Weinberger. 2018. Adversarial deep aver-
aging networks for cross-lingual sentiment classifica-
tion. Transactions of the Association for Computa-
tional Linguistics, 6:557–570.
George Dueñas and Diego Gómez. 2015. A bilin-
gual dictionary with semantic mediawiki: The lan-
guage saliba’s case. The 4th International Confer-
ence on Language Documentation and Conservation
(ICLDC).
Jeff Ens, Mika Hämäläinen, Jack Rueter, and Philippe
Pasquier. 2019. Morphosyntactic disambiguation in
an endangered language setting. In 22nd Nordic Con-
ference on Computational Linguistics (NoDaLiDa).
Linköping University Electronic Press.
Riho Grünthal. 2016. Transitivity in erzya: Second
language speakers in a grammatical focus. Mordvin
languages in the field.
Mika Hämäläinen. 2019. Uralicnlp: An nlp library for
uralic languages. Journal of open source software.
Mika Hämäläinen. 2021. Endangered languages are not
low-resourced! In Multilingual Facilitation. Univer-
sity of Helsinki.
Mika Hämäläinen, Niko Partanen, Jack Rueter, and
Khalid Alnajjar. 2021. Neural morphology dataset
and models for multiple languages, from the large to
the endangered. In Proceedings of the 23rd Nordic
Conference on Computational Linguistics (NoDaL-
iDa), pages 166–177.
Mika Hämäläinen and Jack Rueter. 2019. An open
online dictionary for endangered uralic languages.
Electronic lexicography in the 21st century.
Mika Hämäläinen and Jack Michael Rueter. 2018. Ad-
vances in synchronized xml-mediawiki dictionary
development in the context of endangered uralic lan-
guages. In Proceedings of the XVIII EURALEX In-
ternational Congress. Ljubljana University Press.
Mika Hämäläinen, Liisa Lotta Tarvainen, and Jack
Rueter. 2018. Combining concepts and their transla-
tions from structured dictionaries of uralic minority
languages. In Proceedings of the Eleventh Interna-
tional Conference on Language Resources and Eval-
uation (LREC 2018).
Mika Hämäläinen and Linda Wiechetek. 2020. Mor-
phological disambiguation of south sámi with fsts
and neural networks. In Proceedings of the 1st
Joint Workshop on Spoken Language Technologies
for Under-resourced languages (SLTU) and Collab-
oration and Computing for Under-Resourced Lan-
guages (CCURL), pages 36–40.
Arja Hamari. 2011. The abessive in the permic lan-
guages. Suomalais-Ugrilaisen Seuran Aikakauskirja,
2011(93):37–84.
Nikolaus P Himmelmann. 1998. Documentary and de-
scriptive linguistics. Linguistics, 36(1).
Benjamin Hunt, Emily Chen, Sylvia LR Schreiner, and
Lane Schwartz. 2019. Community lexical access for
an endangered polysynthetic language: An electronic
dictionary for st. lawrence island yupik. In Proceed-
ings of the 2019 Conference of the North American
Chapter of the Association for Computational Lin-
guistics (Demonstrations), pages 122–126.
Mika Hämäläinen and Khalid Alnajjar. 2023. Pyhfst:
A pure python implementation of hfst.Zenodo.
10.5281/zenodo.7791470.
Mika Hämäläinen and Jack Reuter. 2019. Finding
sami cognates with a character-based nmt approach.
In Proceedings of the Workshop on Computational
Methods for Endangered Languages, volume 1.
Ann Irvine and Chris Callison-Burch. 2014. Halluci-
nating phrase translations for low resource mt. In
Proceedings of the Eighteenth Conference on Compu-
tational Natural Language Learning, pages 160–170.
Fred Karlsson, Atro Voutilainen, Juha Heikkilae,
and Arto Anttila. 2011. Constraint Grammar:
a language-independent system for parsing unre-
stricted text, volume 4. Walter de Gruyter.
Gerson Klumpp. 2016. Semantic functions of com-
plementizers in permic languages. Complementizer
Semantics in European Languages, pages 529–586.
Krister Lindén, Erik Axelson, Senka Drobac, Sam Hard-
wick, Juha Kuokkala, Jyrki Niemi, Tommi A Piri-
nen, and Miikka Silfverberg. 2013. Hfst—a system
for creating nlp tools. In Systems and Frameworks
for Computational Morphology: Third International
Workshop, SFCM 2013, Berlin, Germany, September
6, 2013 Proceedings 3, pages 53–71. Springer.
Patrick Littell, Aidan Pine, and Henry Davis. 2017. Wal-
dayu and waldayu mobile: Modern digital dictionary
interfaces for endangered languages. In Proceed-
ings of the 2nd Workshop on the Use of Compu-
tational Methods in the Study of Endangered Lan-
guages, pages 141–150.
Sjur Moshagen, Jack Rueter, Tommi Pirinen, Trond
Trosterud, and Francis M Tyers. 2014. Open-source
infrastructures for collaborative work on under-
resourced languages. Collaboration and Computing
for Under-Resourced Languages in the Linked Open
Data Era, pages 71–77.
Hendry Muljadi, Hideaki Takeda, Shoko Kawamoto,
Satoshi Kobayashi, and Asao Fujiyama. 2006. To-
wards a semantic wiki-based japanese biodictionary.
In SemWiki.
Arbi Haza Nasution, Yohei Murakami, and Toru Ishida.
2018. Designing a collaborative process to create
bilingual dictionaries of indonesian ethnic languages.
In Proceedings of the Eleventh International Confer-
ence on Language Resources and Evaluation (LREC
2018).
Niko Partanen, Rogier Blokland, KyungTae Lim,
Thierry Poibeau, and Michael Rießler. 2018. The first
komi-zyrian universal dependencies treebanks. In
Second Workshop on Universal Dependencies (UDW
2018), November 2018, Brussels, Belgium, pages
126–132.
Jack Rueter, Marília Fernanda Pereira de Freitas, Sid-
ney Da Silva Facundes, Mika Hämäläinen, and Niko
Partanen. 2021. Apurinã Universal Dependencies
treebank. In Proceedings of the First Workshop on
Natural Language Processing for Indigenous Lan-
guages of the Americas, pages 28–33, Online. Asso-
ciation for Computational Linguistics.
Jack Rueter and Mika Hämäläinen. 2017. Synchronized
mediawiki based analyzer dictionary development.
In 3rd International Workshop for Computational
Linguistics of Uralic Languages (IWCLUL 2017).
The Association for Computational Linguistics.
Jack Rueter and Mika Hämäläinen. 2020. Fst mor-
phology for the endangered skolt sami language. In
Proceedings of the 1st Joint Workshop on Spoken Lan-
guage Technologies for Under-resourced languages
(SLTU) and Collaboration and Computing for Under-
Resourced Languages (CCURL), pages 250–257.
Jack Rueter, Mika Hämäläinen, and Khalid Alnajjar.
2023. Modelling the reduplicating Lushootseed mor-
phology with an FST and LSTM. In Proceedings of
the Workshop on Natural Language Processing for
Indigenous Languages of the Americas (Americas-
NLP), pages 40–46, Toronto, Canada. Association
for Computational Linguistics.
Jack Rueter, Mika Hämäläinen, and Niko Partanen.
2020a. Open-source morphology for endangered
mordvinic languages. In Proceedings of Second
Workshop for NLP Open Source Software (NLP-OSS).
The Association for Computational Linguistics.
Jack Rueter, Niko Partanen, and Larisa Ponomareva.
2020b. On the questions in developing computational
infrastructure for komi-permyak. In Proceedings of
the Sixth International Workshop on Computational
Linguistics of Uralic Languages, pages 15–25.
Jack Rueter and Francis Tyers. 2018. Towards an open-
source universal-dependency treebank for erzya. In
Proceedings of the Fourth International Workshop
on Computational Linguistics of Uralic Languages,
pages 106–118.
Pekka Sammallahti and Jouni Mosnikoff. 1991. Suomi-
koltansaame sanakirja: Lää’dd-sää’m sää’nn
ˇ
ke’rjj.
Girjegiisá.
